3 ms·
I think the reason for this gap are the differences in scope and novelty between codebases. When you need an LLM to write a piece of code that's been written a
by SR2Z 2y ago
I think the reason for this gap are the differences in scope and novelty between codebases. When you need an LLM to write a piece of code that's been written a million times before (e.g. "find the normal vector to this plane", "find the highest scoring user") it generally produces decent code.
But on the flip side, this type of code is intrinsically less valuable than novel stuff ("convert this signed distance field to a mesh") which an LLM will choke on.
- throwuxiytayq 2y agoNot sure that vector normalization and “MaxBy” count as a piece of code. It’s a building block, less than a line of code. Usually more compact to just type it out than to describe it in natural language.
- SR2Z 2y agoIt really depends - a recent example I had was trying to implement the DPMO paper (a signed distance field to mesh algorithm), and one of the steps is "compute the plane of best fit with these points and project this other point onto it." Not a particularly long piece of code, but long enough that my local DeepSeek model was able to meaningfully save time for me. http://www.sccg.sk/~chalmo/GM/SM02ob.pdf http://www.sccg.sk/~chalmo/GM/SM02ob.pdf